Re: failuer to build uberbaum --target=m68k-elf

[Jim Wilson <wilson at tuliptree dot org>]

In
	http://gcc.gnu.org/ml/gcc/2003-01/msg01606.html
you mention that combine creates
(and:SF (mem/f:SF (plus:SI (reg/f:SI 14 %a6)
            (const_int 8 [0x8])) [2 x+0 S4 A16])
    (const_int 2147483647 [0x7fffffff]))

The bug here is that we never should have created this rtl because 
AND:SF is not meaningful.  This rtl is created in simplify_comparison 
where it tries to permute AND and SUBREG.  So I believe the correct fix 
is to add a check for an integral mode there, as otherwise this 
optimization is not correct.

I am able to build an uberbaum m68k-elf tree with the following patch, 
except for gdb's lib rda, but that isn't a compiler problem. I will do 
x86 bootstraps tomorrow to verify and then check it into the 3.3 branch 
and the trunk.

2003-02-26  James E Wilson  <wilson at tuliptree dot org>

	* combine.c (simplify_comparison): Require integral mode when
	permuting SUBREG with AND.

Index: combine.c
===================================================================
RCS file: /cvs/uberbaum/gcc/combine.c,v
retrieving revision 1.342
diff -p -r1.342 combine.c
*** combine.c	20 Feb 2003 20:56:51 -0000	1.342
--- combine.c	26 Feb 2003 05:58:53 -0000
*************** simplify_comparison (code, pop0, pop1)
*** 11135,11140 ****
--- 11135,11143 ----
  	     represents the low part, permute the SUBREG and the AND and
  	     try again.  */
  	  if (GET_CODE (XEXP (op0, 0)) == SUBREG
+ 	      /* Require an integral mode, to avoid creating something like
+ 		 (AND:SF ...).  */
+ 	      && SCALAR_INT_MODE_P (GET_MODE (SUBREG_REG (XEXP (op0, 0))))
  	      /* It is unsafe to commute the AND into the SUBREG if the SUBREG
  		 is paradoxical and WORD_REGISTER_OPERATIONS is not defined.
  		 As originally written the upper bits have a defined value